Adder-Subtractor

implement by neural network

Report

Q1. Analyze the results under different number of digits, training epoch, training size ...

Different number of digits

• In 100 epoch comparison, data size 36000

digits	loss	accuracy
3	0.1451441201945146	0.95096875
4	0.5435593464533488	0.8006050003369649
5	0.8493591320435206	0.690370833826065

---

Different training epoch

• Data size 36000

epoch	loss	accuracy
100	0.09084687401056289	0.9685291666666667
200	0.059655092824871345	0.9806083333333333
300	0.05813017793592686	0.98245
400	0.05831938024403838	0.9826479166666666
500	0.05609151270463287	0.9845333333333334

---

Different training size

• In 100 epoch comparison

training size	loss	accuracy
36000	0.0976901263092955	0.96611875
18000	0.3865795685807864	0.85424375
9000	1.160702965593338	0.56805

---

Different training batch size

• Data size 36000

batch size	loss	accuracy
128	0.2030707090501984	0.94389375
256	0.09084687401056289	0.9685291666666667
512	0.34124497500658035	0.8817395833333334
1024	0.927651391617457	0.6435854166666667

Q2. Can we apply the same training approach for multiplication?

• By experiment, it can not apply same training approach on it.

Layer number	loss	accuracy
1	0.7557860101699829	0.7010041667461395
2	0.6974024806499481	0.7185916664600372
3	0.6423751524686814	0.736104166650772